tt-metal models README + model matrix (vendor measured)

Vendor model matrix at tenstorrent/tt-metal models/README.md, re-read 2026-10-16 (page footer says Last Update Dec 8, 2025). ttft = time to first token in ms; t/s/u = 1/inter-token-latency of the first generated token after prefill (per user); t/s = t/s/u * batch, i.e. aggregate across concurrent requests, NOT per-request. Reported LLM numbers use input sequence length 128 except Mamba. Collected from the tt-metal model demos, not the vLLM server path. Starred Blackhole rows are the vendor's own caveat that the software is still under active optimization.